Google acquired artificial intelligent software company, DeepMind, in 2014 for £400 million but has almost certainly already recovered this cost by using the program to manage energy consumption in its data centres. DeepMind had been in discussions with Facebook before these broke down and Google stepped in. At the time, DeepMind had recently demonstrated it’s Alpha Go application defeating a human player at Go. Since then, we have already seen how this has been used within the business: current demonstrations include the software teaching itself to play video games on the old Atari systems, where the system learns the response to making a decision and in effect learns the rules and patterns of the game. Google is also teaching the software how to detect eye diseases. However, whilst being beaten by a sophisticated, multi-million dollar AI system at an Atari game is arguably not the most humbling of experiences, how about if the system could be used to reduce energy consumption? Managing energy consumption is a case of balancing a number of variables, but there are effects to every change so it is a process that software such as DeepMind should be very suitable for. Google certainly thinks so as it has put the DeepMind software in charge of managing its data center energy bills and the software has already made impressive “power efficiency” savings of around 15%.
In the detail, Google operates and manages a number of data centers around the world. These are in effect massive warehouses containing the many servers and storage drives needed to drive Google’s services around the world – which means everything from indexing the Internet, providing Gmail and Google Docs, through to the cloud computing platform that Google is successfully selling to enterprise businesses. Each data center is constructed in a different way, in a different part of the world and so has different requirements and variables. We have already seen how Google is signing up to renewable energy resources around the world, but by letting DeepMind control the lights and heating (and many other factors), the intelligent software has proven itself better able to manage the energy bill. There are somewhere close to 120 variables that the software takes into account when determining how to use the energy, and in cash terms, the savings are considerable. DeepMind changes how the data centere is internally configured by changing how equipment is used, such as opening and closing windows, running or stopping cooling fans, based on its understanding of the environment.
Google uses a “power efficiency” metric to determine efficiency. This is a measure of how much power it is using for its servers compared with how much energy it needs for the associated supporting infrastructure (typically, the cooling system). According to Google, 4.4 megawatt-hours of electricity was used by the whole business in 2014 and a considerable proportion of this would be the data centers it operates. To put this into perspective, 4.4 megawatt hours is more than the energy used to run 350,000 homes. A 15% reduction in this is a not-inconsiderable 0.66 megawatt-hours and each megawatt hour costs from $25 to $40: Google stands to save millions of dollars with the energy savings. Better yet, because DeepMind learns from experience, Google expects the energy savings eked by the software to improve over time and DeepMind may be able to double the already substantial energy savings achieved. This is likely to tip the cost savings into the billions of dollars per year figure. Google spent £400 million on DeepMind, which in 2014 was around $600 million, and it is likely that the software has already paid for itself.
Going forwards, given that the more information DeepMind has about its environment, the better the system is able to learn and refine its control software. DeepMind’s engineers are learning where the software has less information and it is likely that they will be asking Google to incorporate more sensors into its data centers in order to better optimize the power consumption. DeepMind’s ability to manage energy usage is likely to be an important cornerstone technology in tomorrow’s smarter cities: as mankind becomes more aware of our impact on the planet and as energy becomes more and more expensive, and as the technology becomes more responsive, it is relatively easy to see how smart home thermostats linked with smart power stations and weather sensors could reduce total energy bills by a considerable amount.